docs(prds): design the code graph as an MCP surface

Supersedes symbol-reference-index.md, which scoped only the data layer and
left the interface unstated. Deleted rather than archived: it was written
today, never implemented, and keeping both would leave two documents claiming
authority over one capability.

Three things the narrower doc did not settle.

The interface is MCP. archied consumes MCP and serves none, so exposing the
graph that way reaches Claude Code, pi, codex and Archie's own agents through
one surface with no coupling to the daemon. Diagrams are deliberately not a
tool: a model handed an accurate subgraph writes its own Mermaid, and the
valuable half was never the rendering.

Every answer declares what it could not see. RTA over-approximates interface
dispatch and reflection is invisible, so the *extract yaegi tables make
symbols look reachable regardless. An agent trusting a confident wrong graph
is worse off than one that greps.

A code graph does not power a documentation chatbot. It answers structural
questions about Go symbols; answering "how do I configure a second identity"
means retrieving prose from markdown. I had previously implied one substrate
served both. It does not, and building it that way would serve neither well.
This commit is contained in:
sam committed 2026-09-18 13:42:33 +10:00
1 parent 2c805ef15d
commit 98c4032285
2 files changed
+110 -117

No files matched your search

+110
View File
@@ -0,0 +1,110 @@
# Code graph over MCP (`tools/codegraph`)
Status: proposed. Supersedes `symbol-reference-index.md`, which scoped only the
data layer.
## The problem
Establishing what a change touches currently costs an agent eight to fifteen
greps, and the answers are wrong in specific ways. `grep -rn "Register("`
cannot separate `plugin.Registry.Register` from `captureintake.Receiver.Register`
from `forgerpc.Server.Register`, all of which exist here. It also matches the
sibling checkouts under `.worktrees/`, which are near-complete copies of the
module, so every count is silently inflated.
That cost lands hardest in investigation, triage and the TDD loop, where the
first question is always "what does this touch".
## Decision: exact analysis, not semantic retrieval
Tools like Greptile parse with tree-sitter, summarise each node, embed the
summaries, and retrieve semantically. That design exists because they must
serve any language. This repository is one Go module, and Go's type checker is
ground truth:
| Question | Source of truth |
| --- | --- |
| Who references this symbol? | `go/packages` + `go/types`, resolved by identity |
| What calls what? | `callgraph` RTA over SSA |
| What imports what? | `go list -deps` |
No vector database, no embedding cost, no edges invented by a summary that
drifted from the code. For a single-language repository this is both cheaper
and strictly more accurate. Copy the interface those tools expose, not their
architecture.
**A code graph does not power a documentation chatbot.** It answers structural
questions about Go symbols. Answering "how do I configure a second identity"
means retrieving prose from markdown, which needs chunking, embedding and
similarity search over a different corpus entirely. That remains a separate
capability with its own design. Conflating the two would build one system that
does neither well.
## Interface: an MCP server
`archied` already consumes MCP as a client (`internal/app/agentworker`
starts providers) and serves none. Exposing the graph over MCP reaches Claude
Code, pi, codex and Archie's own agents through one surface, with no coupling
to the daemon: it is configured like any other entry in `[[mcp_servers]]`.
Four tools, each returning JSON with file:line citations:
- `references(symbol)` — every reference site, classified `prod` / `test` /
`generated` / `dynamic`
- `callers(symbol, depth)` — inbound call edges
- `callees(symbol, depth)` — outbound call edges
- `blast_radius(file | symbol)` — the transitive neighbourhood a change reaches,
which is the single call meant to replace the grep sequence
Diagrams are deliberately not a tool. A model given an accurate subgraph writes
its own Mermaid; a diagram endpoint would only fix the rendering of something
the model already has.
## Decisions
**Every answer declares what it could not see.** RTA over-approximates
interface dispatch, and reflection is invisible: every `reflect.ValueOf` entry
in `wfextract`, `gateextract`, `pluginextract`, `secretextract` and
`logextract` makes a symbol look reachable whether or not anything calls it.
Each response carries a `dynamic_edges` field naming those boundaries. An agent
trusting a confident wrong graph is worse off than one that greps, so the graph
states its own limits rather than implying completeness.
**Warm in memory, keyed on HEAD, never persisted to disk.** `deadcode` does
whole-program RTA in under seven seconds; this does less work. The server
rebuilds when HEAD changes and refuses its cache outright when the tree is
dirty. Nothing is written to disk, because a persisted index becomes a rendered
view that outlives its evidence. That is the failure that made `STATUS.md`
report 90 of 91 findings re-verified while most statuses were 77 commits stale,
and the same one that made `reachaudit` serve a previous tree's analysis from
`/tmp` until its cache was keyed on HEAD.
**`tools/codegraph`, not `internal/`.** It is developer tooling, not daemon
runtime, so it belongs beside `docsgen`, `contractaudit` and `reachaudit` in
the `tools` module, which `task test:tools` already covers. Start as a single
Go file; split only when it earns it.
**No new dependency.** `golang.org/x/tools` is already in `go.sum` at v0.49.0.
## How it is verified
Red first. Table-driven tests against small fixture modules built in
`t.TempDir()`, asserting the behaviour that distinguishes this from grep:
1. Two methods spelled `Register` on different types resolve to two separate
records, and a reference to one does not increment the other.
2. A symbol referenced only from `_test.go` reports zero `prod` references.
3. A symbol referenced only from a `reflect.ValueOf` table reports zero `prod`
references and one `dynamic`, and names the table in `dynamic_edges`.
4. A generated file's references land in `generated`, not `prod`.
Mutation-verify each: reintroduce the conflated behaviour and watch the
assertion fail.
Acceptance is a live cross-check rather than a unit test. Run it against this
tree and confirm it reports zero production references for
`worktree.Manager.Resume`, `webui.NewDangerousService` and
`secret.DefaultRegistry`. All three were confirmed dead by hand, and `deadcode`
either stayed silent on them or, in the case of a package var, cannot represent
them at all. Agreeing with `deadcode` is not the bar. Catching what `deadcode`
structurally cannot is.
-117
View File
@@ -1,117 +0,0 @@
# Type-resolved symbol reference index (`tools/symbolrefs`)
Status: proposed. Single-file Go tool in the `tools` module.
## The gap
Three questions get asked constantly during audits, and each currently gets
answered by a different unreliable method:
| Question | Today | Problem |
| --- | --- | --- |
| Is this reachable from a shipped binary? | `task deadcode` | Sound but incomplete, functions only |
| Who references this symbol? | `grep` / `ast-grep` | Matches spelling, not identity |
| Is this package in any binary? | `tools/reachaudit` | Package granularity |
The middle row is the hole. `grep -rn "Register("` cannot tell
`plugin.Registry.Register` from `captureintake.Receiver.Register` from
`forgerpc.Server.Register`, all of which exist here. It also matches the three
sibling checkouts under `.worktrees/`, which are near-complete copies of the
module, so every count is inflated by a factor nobody notices.
This matters because reachability alone does not settle a finding. RTA
over-approximates: it assumes every address-taken function may be called, and
every `reflect.ValueOf` entry in `wfextract`, `gateextract`, `pluginextract`,
`secretextract` and `logextract` is address-taken. `deadcode` is therefore
silent on `worktree.Manager.Resume` despite it having zero callers anywhere in
the tree. `deadcode` also reports functions only, so a dead package var like
`secret.DefaultRegistry` is structurally invisible to it.
A reference index answers what reachability cannot: **a symbol with zero
non-test reference sites is dead regardless of what RTA believes about its
address.** The two reports are complementary, and neither subsumes the other.
## What it does
`packages.Load` over the root module with `NeedTypes | NeedTypesInfo |
NeedSyntax | NeedDeps`, then for every identifier in every file resolve
`TypesInfo.Uses[ident]` to a `types.Object` and attribute the reference to its
declaring object. Identity comes from the type checker, so spelling collisions
and sibling worktrees both disappear: `go list` resolves by import path and
never walks `.worktrees/`.
Each reference site is classified by the file that holds it:
- `prod` — an ordinary non-test file in the module
- `test` — `_test.go`
- `generated` — matches the standard generated-file header, plus
`internal/contracts/**` and `ui/dist/`
- `dynamic` — a `reflect.ValueOf` table or other runtime-loading site, which
means static analysis cannot see the real caller
`dynamic` is a separate class rather than folded into `prod` because the whole
point is to stop those sites laundering a dead symbol into a live-looking one.
Output is JSONL on stdout, one record per package-level declaration:
```json
{"symbol":"internal/worktree.Manager.Resume","kind":"method",
"decl":{"file":"internal/worktree/worktree.go","line":261},
"refs":{"prod":0,"test":4,"generated":0,"dynamic":0},
"sites":[{"file":"internal/worktree/worktree_test.go","line":88,"class":"test"}]}
```
So the query that took me six hand-run greps this session becomes one line:
```sh
task symbolrefs | jq -c 'select(.refs.prod == 0 and .refs.dynamic == 0)'
```
## Scope
In: package-level funcs, methods, types, vars, consts in the root module.
Out: local variables, struct fields, reachability (that is `task deadcode`),
and any write path. The tool is read-only and advisory. It never proposes
deletion, matching `reachaudit`'s existing stance.
## Decisions
**Single file, `tools/symbolrefs/symbolrefs.go`.** It is dev tooling, not
product code, so it belongs beside `docsgen` and `contractaudit` in the `tools`
module rather than under `internal/`. That module is already covered by
`task test:tools`, which `task check` runs.
**No persisted index.** `deadcode` does whole-program RTA in under 7 seconds;
this does strictly less work (no SSA, no call graph). Anything cached becomes a
rendered view that outlives its evidence, which is the exact failure that made
`STATUS.md` claim 90 of 91 findings re-verified while most statuses were 77
commits stale, and the same reason `reachaudit`'s `/tmp` cache had to be keyed
on HEAD. Recompute; do not reconcile.
**Not in `task check`.** It reports a standing backlog, so gating on it fails
every run until the backlog is zero. Same rationale as `task deadcode` and
`task vuln`.
**No new dependency.** `golang.org/x/tools` is already in `go.sum` at v0.49.0.
## How it is verified
Red first, per the repo protocol. The tests are table-driven against small
fixture modules built in `t.TempDir()`, asserting the classification that
distinguishes this tool from grep:
1. Two distinct methods spelled `Register` on different types resolve to two
separate records, and a reference to one does not increment the other.
2. A symbol referenced only from `_test.go` reports `prod: 0`.
3. A symbol referenced only from a `reflect.ValueOf` table reports `prod: 0`
and `dynamic: 1`, not `prod: 1`.
4. A generated file's references land in `generated`, not `prod`.
Mutation-verify each: reintroduce the conflated behaviour and watch the
assertion fail, per `HANDOFF.md` rule 5.
Acceptance is a live cross-check, not a unit test: run it against the tree and
confirm it reports `prod: 0` for `worktree.Manager.Resume`,
`webui.NewDangerousService` and `secret.DefaultRegistry`, all three of which
were confirmed dead by hand this session and all three of which `deadcode`
either missed or could not represent.